PaperlayerView exact revision on arXiv

Source-linked AI summary

LACRM: LLM-Guided Mining of Directed Advertiser Competition Relationships in E-Commerce Advertising

Xiang Shi, Yitian Kou, Bing Li, Shengsheng Niu, Chun Gan, Linghui Meng, Jie He, Zhangang Lin, Ching Law

arXiv:2610.13959v1cs.LG

TL;DR

Advertiser competition is directed, multi-faceted, and only partly visible in auction histories, despite its relevance to advertiser performance. LACRM combines role-separated profiling, asymmetric multi-route retrieval, and behavioral ranking to model directed competition and retrieve latent rivals. It outperforms the evaluated directed GNN baselines and ablated variants, with assessments supporting the relevance of recovered latent competitors.

  • Problem

    Advertiser-facing systems generally infer competition indirectly from auction outcomes, although advertiser performance depends on rivals and competition is directed, multi-faceted, and partially observed.

    Source: §1 · p. 1

  • Method

    LACRM constructs a directed, weighted competition graph using role-separated advertiser and query profiles, asymmetric multi-route retrieval, and listwise ranking aligned with auction behavior.

    Source: §1 · p. 2

  • Results

    LACRM outperforms the evaluated directed GNN baselines and ablated variants in realized-rival retrieval and ranking, while independent LLM and human assessments support the relevance of latent competitors.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    The framework provides an explicit, reusable competition layer for advertiser-facing optimization tasks.

    Source: §1 · p. 2

    Historical ground truth reflects only realized auction clashes, so an unobserved latent substitute may be counted as a false positive under Greal.

    Source: §4.1.2 · p. 5

Abstract

from arXiv · show

In e-commerce advertising, ad platforms typically fail to model the externalities among advertisers when making optimization decisions: an advertiser's actual performance is largely shaped by rivals competing for the same users; these rivals typically operate in overlapping product categories and at comparable budget scales. However, capturing this rivalry as an explicit, reusable signal is challenging because it is directed, multi-faceted, and partially observable. Existing paradigms such as interaction graphs, taxonomy rules, and single-embedding models do not fully capture these properties. In this paper, we propose LACRM (LLM-driven Advertiser Competition Relationship Mining), a novel framework that constructs directed competition graphs with continuous edge intensities via a three-stage pipeline. LACRM moves beyond symmetric similarity through LLM-driven role-separated profiling, which enables asymmetric multi-route retrieval to uncover both realized and latent rivals. The framework further incorporates listwise learning-to-rank aligned with temporal auction behavior to accurately model directed competition intensity. In offline retrieval experiments on an industrial benchmark, LACRM outperforms strong directed GNN baselines and ablation variants. Multi-LLM and human studies provide complementary evidence that the recovered latent rivals reflect meaningful market competition rather than merely model-generated artifacts. In production A/B tests, LACRM-guided bidding and incentive allocation increase advertising revenue by 1.73% and 1.14%, respectively. The results demonstrate the practical value of explicit competition modeling in large-scale advertising systems.

1 Introduction

Advertiser performance depends on rivals, yet current systems generally infer competition indirectly from auction outcomes rather than representing it as a reusable signal. LACRM models directed competition and retrieves latent rivals, with evaluations reporting stronger retrieval and ranking than the tested baselines and ablations.

1 Introduction

  • Advertiser performance depends on competitors targeting overlapping users, products, and budgets, but existing systems typically infer rivalry indirectly from auction outcomes.The systems instead rely primarily on each advertiser’s attributes, campaign history, and bidding outcomes.Source: §1 · p. 1
  • Competition is directed, spans multiple facets, and is only partially observed, leaving latent rivals—especially new and long-tail advertisers—without direct auction evidence.The passage reports that 99.24% of observed competition edges show asymmetry in direction or intensity.Source: §1 · p. 1
  • Taxonomy rules lack fine-grained strength, behavioral graph models are limited by observed interactions, and shared text embeddings conflate advertiser and competitor roles.The paper also notes that dedicated competitor-mining methods commonly model symmetric, undirected, or binary relations.Source: §1 · pp. 1–2
  • LACRM combines role-separated profiles, multi-route semantic retrieval, and listwise behavioral alignment to capture directed competition and recover latent rivals beyond observed auction graphs.Its competition layer is intended as a reusable market representation for advertiser-facing optimization tasks.Source: §1 · p. 2
  • Evaluation covers realized-competitor retrieval, assessments of latent competitors, and production applications; LACRM beats the tested directed GNN baselines and ablations, with deployment gains in revenue and budget fulfillment.The passage does not give numerical values for these production gains.Source: §1 · p. 2

2 Problem Formulation

The paper formulates advertiser competition as learning a directed, weighted graph whose edge weights represent asymmetric competitive intensity. Auction co-occurrence provides evidence of realized rivalry, but the graph-learning objective also calls for discovering latent competitors.

2 Problem Formulation

  • The target is a directed, weighted graph in which an edge from advertiser v_i to v_j records v_i’s view of v_j as a competitor, with a directional intensity score.The learned score matrix is intended as a foundational market-structure representation for downstream tasks.Source: §2 · p. 2
  • Competition intensity is operationally approximated by traffic diversion in auction co-occurrences, while latent discovery addresses rivals absent from historical interactions.Auction co-occurrence is described as reliable but incomplete evidence of realized rivalry.Source: §2 · p. 2
  • Only 0.76% of observed edges are fully symmetric in both directions and intensity; 99.24% are asymmetric in at least one respect.The corpus contains over 70 million directed competition edges.Source: §2 · p. 2

3 Methodology

LACRM turns merchant profiles into directed competition scores through dual-profile generation, asymmetric multi-route retrieval, and temporal behavioral reranking. Its stages separate advertiser identity from rival-seeking intent, retrieve candidates by competition view, and align their scores with auction behavior.

3.1 LLM-Guided Dual-Profile Generation

  • Advertiser Representation captures a merchant’s invariant identity, while multi-view Query Representations express the rival cohorts it targets.Both roles use an isomorphic six-field schema covering category and brand, search intent, style and positioning, audience and scenario, price band, and shop identity.Source: §3.1 · p. 2

3 Methodology

  • The three-stage pipeline generates dual profiles, retrieves candidates through multiple routes, and reranks them with temporal behavioral signals to produce directed scores.The stages are Dual-Profile Generation, Multi-Route Retrieval, and Dynamic Behavioral Alignment.Source: §3 · p. 2, Figure 1 · p. 3

3.1 LLM-Guided Dual-Profile Generation

  • Five competition views separate distinct rivalry patterns; each view uses a dedicated route, and prior realized rivals can seed generation without restricting it to observed pairs.For cold-start advertisers, generation instead relies on the merchant’s own attributes.Source: §3.1 · p. 3

3.2 Asymmetric Embedding and Multi-Route Retrieval

  • Role-separated inputs and task-specific instructions produce asymmetric embeddings, with an independent retrieval route for each view and view-specific scores retained as candidate features.The routes retrieve nearest neighbors for each view before merging candidates at the directed merchant-pair level.Source: §3.2 · p. 4

3.3 Dynamic Behavioral Alignment

  • LambdaMART reranks each source advertiser’s candidate list using view scores, hit indicators, and temporal auction features, then normalizes scores per source to [0, 1].In cold-start cases, candidates can be scored from semantic and static features without behavioral features.Source: §3.3 · p. 4

4 Experiments

The experiments evaluate LACRM on industrial advertising data through realized-rival retrieval, latent-rival verification, and production applications. They report stronger offline retrieval than directed GNN baselines and ablations, alongside evidence that discovered latent rivals reflect meaningful competition.

4.1.1 Dataset and Realized Ground Truth.

  • The evaluation uses multi-week ad-serving logs and defines directed realized-rival edges when traffic diversion exceeds 0.05 during an observation window.The evaluation window is isolated from model fitting, with ranker labels computed one week earlier to prevent target leakage.Source: §4.1.1 · pp. 4–5
  • Over 40% of Top-100 realized rivalries are newly emerging edges absent from the preceding four-week graph.Novelty fractions are 2.7%, 23.6%, and 40.5% at K∈{10, 50, 100}.Source: §4.1.1 · p. 5

4.1.2 Evaluation Metrics.

  • Precision@K, Recall@K, and NDCG are averaged across source advertisers with realized rivals, using historical auction co-occurrence as continuous relevance grades.The same evaluation set and uniformly applied directional decoding are used for all systems.Source: §4.1.2 · p. 5

4.1.3 GNN Baselines.

  • The GNN references share LACRM’s directed graph, behavioral logs, node universe, and metric protocol, while varying their node-level inputs.GNN (External) uses behavioral temporal features; GNN (w/ Qwen Emb.) additionally uses the pretrained Qwen embeddings.Source: §4.1.3 · p. 5

4.1.4 Implementation

  • Profiles are regenerated weekly, while behavioral features are recomputed daily to reflect different rates of merchant-identity and auction-dynamics change.The ranker is a LambdaMART ensemble with a listwise NDCG objective.Source: §4.1.4 · p. 5

4.2.1 Comparison with GNN Baselines.

  • LACRM (Full) outperforms GNN (External) on every metric and consistently achieves the best results across all evaluation metrics.The same Qwen embeddings recover part of the gap, but GNN (w/ Qwen Emb.) still trails LACRM substantially in recall.Source: §4.2.1 · p. 5, §4.2.2 · p. 5, Table 1 · p. 6

4.2.2 Ablation Study.

  • The ablation finds that LACRM (Full) performs best across metrics, with supervised profile synthesis contributing more than the smaller complementary GRPO refinement.Variants successively remove profile alignment, multi-view retrieval, post-training, or GRPO while retaining specified portions of the pipeline.Source: §4.2.2 · p. 5

4.3.1 LLM-Judge Verification.

  • All three LLM judges rank realized rivals above latent pairs and latent pairs above negative controls; the latent cohort has a 61.3% ensemble positive rate.The latent cohort is significantly above negative controls and remains distinct from realized rivals.Source: §4.3.1 · p. 6

4.4 Coverage and Generalization Analysis

This section examines retrieval-route activation, recovery of emerging competition edges, and production effects of LACRM-derived signals. Route activation is uneven, emerging-edge recall rises at larger cutoffs, and both production experiments report gains.

4.4.1 Per-View Activation.

  • Dense routes account for 80–88% of top-ranked candidates, while specialized sparse routes expand coverage to less frequent patterns at 17–37%.Activation is measured as the fraction of Top-K candidates retrieved across K∈{10, 50, 100}.Source: §4.4.1 · p. 6

4.4.2 Generalization to Emerging Competition Edges.

  • At K={10, 50, 100}, emerging-edge Recall is 8.6%, 77.9%, and 86.0%; these edges were inactive during the preceding four weeks.The recall gap between historical and emerging edges shrank from 66.80 percentage points at K=10 to 16.39 at K=50 and 8.41 at K=100.Source: §4.4.2 · p. 7

4.5 Online Experiments

  • Both 14-day production experiments randomized 200 graph-derived advertiser clusters 100:100 and estimated treatment effects using difference-in-differences.The partition was held fixed across arms, with cluster-level standard errors and randomization-inference significance checks.Source: §4.5 · pp. 7–8

4.5.1 Competition-Aware Bidding Experiment.

  • LACRM-guided bidding increased ad revenue 1.73% and daily budget-target fulfillment 0.91 percentage points.Both effects were statistically significant under the cluster-level inference procedure.Source: §4.5.1 · p. 8

4.5.2 Competition-Aware Incentive Experiment.

  • LACRM-guided incentives increased ad revenue 1.14% relative to the existing incentive policy.The effect was statistically significant under the cluster-level inference procedure (p < 0.01).Source: §4.5.2 · p. 8

5 Related Work

Prior work studies competitive externalities, asymmetric market structures, and semantic retrieval, but generally does not operationalize directed, continuously weighted competition at the advertiser level. LACRM applies role-separated retrieval and auction-behavior alignment to that problem.

5 Related Work

  • Prior advertiser and market-structure methods address competitive spillovers or asymmetric brand effects, while LACRM focuses on advertiser-level directed, continuous competition aligned with auction behavior.The passage relates auction-derived traffic diversion to diversion-ratio formulations in antitrust analysis.Source: §5 · p. 8
  • Product- and firm-level mining captures substitutability, similarity, or observed graph relations, not continuously varying directional competitive pressure among advertisers.The passage describes behavioral methods using co-purchase and basket data [9] [28] and graph methods using company networks or financial and inter-firm knowledge graphs [7] [12] [36].Source: §5 · p. 8
  • General semantic retrieval and LLM-based recommendation target relevance or user–item ranking, whereas LACRM combines role-separated retrieval with temporal auction-behavior alignment.The related-work passage names E5, BGE, and Qwen3-Embedding as semantic retrieval encoders [41] [46] [49].Source: §5 · p. 8

6 Conclusion

LACRM mines directed, weighted advertiser competition through role-separated profiling, asymmetric retrieval, and listwise behavioral alignment. The paper reports retrieval gains, support for latent-rival relevance, and positive production outcomes.

6 Conclusion

  • LACRM outperforms evaluated directed GNN baselines and ablated variants in realized-rival retrieval and ranking, while independent LLM and human assessments support latent-rival relevance.The conclusion reports production gains of 1.73% in ad revenue and 0.91 percentage points in budget-target fulfillment for bidding, plus 1.14% in ad revenue for incentives.Source: §6 · p. 8
  • Production experiments report a 1.73% ad-revenue increase and 0.91 percentage-point budget-target fulfillment improvement for bidding, plus a 1.14% ad-revenue increase for incentives.Source: §6 · p. 8
  • The framework integrates role-separated advertiser profiling, asymmetric multi-route retrieval, and listwise behavioral alignment to capture heterogeneous and directional pressure and discover latent rivals.Source: §6 · p. 8

A Query-View and Behavioral-Feature Specifications

The specification covers the field subsets and competitive boundaries of five query routes, as well as the behavioral profile constructed for the ranker.

A Query-View and Behavioral-Feature Specifications

A.1 Five-View Field Assignment

Five query routes identify distinct forms of advertiser rivalry, from interchangeable products to brand and channel relationships, cross-category substitutes, and time-bound campaign clashes.

A.1 Five-View Field Assignment

  • Direct Homogeneous Rivalry provides a high-recall structural anchor by comparing interchangeable commodities pairwise.It uses fields f1, f2, and f5.Source: §A.1 · p. 9
  • Intra-Brand Structural Expansion retrieves one brand across divergent L3 categories, while Intra-Brand Cross-Channel Friction captures the same brand and category across channels.Both use fields f1 and f6 and are separated by a prompt-level boundary gate.Source: §A.1 · p. 9
  • Cross-Category Functional Substitution masks category to find implicit substitutes competing for the same budget or usage scenario, including latent long-tail threats.This route uses fields f2, f3, f4, and f5.Source: §A.1 · p. 9
  • Temporal Campaign Clashes retrieves rivals during promotional windows and uses fields f2 and f4.Without festival context, its query is inactive, skipped, and contributes zero hit indicators; active requests use the shared index.Source: §A.1 · p. 9

A.2 Behavioral Feature Construction

The behavioral profile combines traffic, conversion, market dynamics, lifecycle, quality, and category-relative measures across six feature groups.

A.2 Behavioral Feature Construction

  • Traffic Scale tracks log-scaled impressions, clicks, and GMV over rolling 1-, 7-, and 14-day windows, while Conversion Efficiency uses short-window CTR, CVR, and average order value.Source: §A.2 · p. 9
  • Volumetric Dynamics includes market-share trajectories, tier-wise Z-scores, variability, and traffic-saturation measures; Ecosystem Rhythm & Lifecycle captures periodicity, active ad-serving days, and platform tenure.Source: §A.2 · p. 9
  • Behavioral Quality measures funnel and promotion signals, while Intra-Category Quantiles & Trends uses category-bounded GMV percentiles and multi-lag autocorrelations.The latter tracks competitive displacement within the same vertical.Source: §A.2 · p. 9

B GNN Reference Implementation Details

The GNN references share a directed graph and training protocol but vary in node inputs; their relation-aware encoder and asymmetric decoder model directional competition.

B GNN Reference Implementation Details

  • The two GNN references share LACRM's graph, behavioral logs, node universe, directed ground truth, and Top-K retrieval protocol, enabling diagnostic comparisons of LLM-guided design.Source: B GNN Reference Implementation Details · p. 9
  • The graph uses realized competition records strictly before the label window, preventing temporal leakage, and represents each rivalry with typed forward and reverse edges.Separate relation channels let the encoder propagate incoming and outgoing competitive pressure.Source: B GNN Reference Implementation Details · p. 9
  • Daily observations for a directed pair are aggregated into one edge weight combining realized competition intensity with temporal decay.The decay constant is set to 14 days.Source: B GNN Reference Implementation Details · p. 9
  • The references share an encoder, decoder, and objective but differ in node inputs: GNN (External) uses behavioral features, while GNN (w/ Qwen Emb.) uses pretrained Qwen embeddings of advertiser profiles.The External baseline omits LACRM's LLM-derived profiles, semantic score, and semantic alignment objective.Source: B GNN Reference Implementation Details · p. 9
  • A two-layer directed, relation-aware Graph-SAGE encoder uses separate incoming and outgoing transformations, while untied source- and target-role heads produce asymmetric edge scores.The encoder weights neighborhood messages by edge weight and concatenates them with a root transform; the decoder combines role vectors and bilinear interaction features.Source: B GNN Reference Implementation Details · p. 9
  • Both references use LACRM's listwise LambdaRank objective and are trained to convergence with validation-NDCG early stopping.All runs stopped before the 10-epoch cap.Source: B GNN Reference Implementation Details · p. 9
Loading 2610.13959v1…